System Prompts Leaks GitHub Walkthrough: ChatGPT, Claude, Gemini & Guardrail Analysis

The asgeirtj/system_prompts_leaks repository is a massive open-source collection documenting hidden system prompt instructions from every major AI chatbot: Claude, ChatGPT, Gemini, Grok, Copilot, Perplexity, and more.

Recent additions include prompts for Kimi K2.6, GPT-5.6, Claude Sonnet 5, and Claude Design (Opus 4.8 with 48 tools and 16 skills).

System prompts are hidden instructions injected before every conversation — they define personality, capabilities, safety boundaries, and tool access. Normally, they are completely invisible to the end user.

This repo provides:

• Transparency — see the exact rules and biases shaping AI responses
• Research — study safety approaches across providers
• Comparison — side-by-side provider configuration analysis
• Education — learn from production-grade system prompts
• Accountability — explain unexpected or surprising AI behavior

Anthropic / Claude: Fable 5, Opus 4.8, Sonnet 5, Claude Code, Cowork, Desktop, Design, Mobile iOS, Chrome extension, Office plugins. Includes raw prompts + tool definitions + skills + diffs between releases.

OpenAI / ChatGPT: GPT-5.6, 5.5 (Thinking/Instant/API), 5.4, 5.3, Codex CLI. Tools: web search, deep research, Python, Canvas, image gen, memory. 8 personality variants (Friendly, Professional, Candid, Cynical, Nerdy, Quirky).

Google / Gemini: 3.5 Flash, 3.1 Pro, Gemini CLI, Antigravity, Jules, NotebookLM.

How each company implements guardrails differently:

• Safety policies: OpenAI has explicit image safety + automation context. Claude has detailed refusal patterns + ethics guidelines.
• Tool access controls: Claude Design = 48 tools + 16 skills. Codex = sandboxed execution environment.
• Personality constraints: GPT-5.1 has 8 distinct personality modes. Grok has dedicated personas.
• Memory systems: OpenAI memory tool prompts, Anthropic reminders system.
• Diff analysis: DiffChecker links between versions to track guardrail evolution over time.

All prompts are stored as individual Markdown files organized by provider directory. The flat structure means no special tooling is needed.

• Clone the repo + use grep or ripgrep to search across all prompts simultaneously
• Track changes via git history
• Compare how different models handle the same capability
• Flat Markdown structure = zero dependencies for browsing

Getting started is as simple as it gets — pure documentation, no build steps:

1. git clone https://github.com/asgeirtj/system_prompts_leaks.git
2. Browse the provider directories
3. No dependencies, no build steps, no configuration

The README serves as a comprehensive index with a "Recently Updated" table showing the latest additions. Licensed under CC0-1.0 — fully open, no restrictions. PRs welcome for new prompts.

Publishing system prompts has a dual-use nature:

• Prompt injection: Published prompts enable studying guardrail weaknesses
• Jailbreaking: Easier when exact safety instructions are known
• Competitive intelligence: Reveals proprietary implementation details
• Arms race: Companies increasingly add obfuscation in response
• CC0 license: Anyone can use the content for any purpose
• Versioning risk: Published prompts may already be outdated

Limitations:

• Point-in-time snapshots only — prompts change frequently
• Extraction may be incomplete
• Missing runtime context (A/B testing, server-side logic)
• No API or structured query interface
• Community maintenance burden
• Verification challenges — hard to confirm completeness